Last updated: August 14, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
AI is threatening natural resources for billions of people mainly through three channels: electricity, water, and land, all three of which are projected to grow faster than AI itself gets more efficient. A UN University report published June 3, 2026 put numbers on it for the first time at global scale, and the totals are large enough to matter to anyone who pays a utility bill, not just data center operators.
Short answer: A June 2026 UN University (UNU-INWEH) report projects AI-related data centers will use 945 terawatt-hours of electricity, 9.3 trillion liters of water, and 14,500 square kilometers of land by 2030. The water figure alone equals the basic annual domestic water needs of the 1.3 billion people living in Sub-Saharan Africa. Efficiency gains per query are real but are being outpaced by total AI usage growth.
I cover AI infrastructure spending and utility data for this site every month, and when I dug into the underlying UNU-INWEH report instead of just the headline, the most surprising number wasn't the electricity — it was how much of the footprint comes from inference, the everyday act of running a chatbot, not from training giant models. That changes who's actually responsible for the number: it's less "the labs" and more the collective habit of 2.5 billion daily ChatGPT prompts, plus every other model's traffic on top.
What the UN report actually found
The report, *Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints*, comes from the United Nations University Institute for Water, Environment and Health (UNU-INWEH) and was released June 3, 2026. Its authors, led by Prof. Kaveh Madani, argue that carbon-only accounting hides the real cost of AI. Low-carbon is not automatically low-water or low-land. A data center can be efficient on emissions and still be a heavy water draw if it's cooled with evaporative systems in a dry region.
By 2030, the report projects global AI-related data center electricity use will hit 945 terawatt-hours. That's nearly triple the combined annual electricity use of Pakistan, Bangladesh, and Nigeria — three countries home to more than 650 million people. It's also roughly double France's entire 2025 consumption, and about 3% of projected world electricity use. The water footprint reaches 9.3 trillion liters, equal to the basic annual domestic water needs of the 1.3 billion people living in Sub-Saharan Africa. Land use hits 14,500 square kilometers, about twice the Jakarta metro area. Carbon emissions reach 399 million tonnes. The report says that would take roughly 6.7 billion trees a decade to offset — about twice the number of trees estimated to exist in the UK.
The part that surprised me most: 80-90% of that total energy use comes from inference, not training. That's the everyday act of running a chatbot, not the one-time cost of building it. ChatGPT alone processes an estimated 2.5 billion prompts a day. On OpenAI's own numbers cited in the report, that adds up to roughly 383 gigawatt-hours a year. The authors also flag that AI compute is extremely concentrated. Over 90% sits in just two countries, the US and China, while more than 150 countries have no sovereign AI compute infrastructure at all. Whatever resource strain AI creates, it's not evenly distributed, and neither is the benefit.
How much water does one AI query actually use?
This is where public numbers get messy. I'd rather be upfront about that than pick the most dramatic figure. Estimates for a single AI query's water footprint range by more than 1,500x. It depends on who measured it, which model, and what counted as "water use."
| Source | Estimate | What it covers | Date |
|---|---|---|---|
| Google (Gemini, official disclosure) | 0.26 mL per median text prompt | On-site cooling + data center cycle, measured | Aug 2025 |
| Sam Altman / OpenAI | ~0.32 mL per average ChatGPT query | Self-reported, methodology not published | 2025 |
| UC Riverside (Shaolei Ren, revised) | ~15 mL per GPT-4-class prompt | On-site cooling plus power-plant water use | 2025 |
| Washington Post / UC Riverside (original) | 519 mL per 100-word GPT-4 email | Widely cited but now considered outdated | 2024 |
| "How Hungry is AI?" benchmark study | <2 mL (efficient models) to 150+ mL (reasoning models) | Cross-model benchmark, varies heavily by task | 2025 |
In my testing of how these numbers get used in headlines, the 519 mL figure from 2024 is still the one that circulates most on social media. Both the original researcher and newer, more granular measurements have moved well below it since, per Forbes’ July 2026 reporting. Google's own disclosure — the only one that's independently auditable rather than self-reported — puts a routine text prompt closer to five drops of water. The water use spikes with reasoning models and image or video generation. The UN report measured a single AI image at 1,450x the energy of a basic text classification, and a single AI video at 200,000x.
Step-by-step: check and shrink your own AI footprint
You don't need to stop using AI tools to act on any of this — you need to know which parts of your usage are actually resource-heavy and which aren't.
1. Separate your "quick text" use from your "heavy generation" use
Chat-based Q&A, summarizing, and short drafting are the cheapest tasks by a wide margin. Image generation, video generation, and long reasoning-mode queries are where the multiplier jumps — sometimes by three orders of magnitude per the UN report's task-level comparisons above.
2. Check whether your provider publishes real numbers
As of this writing, Google is the only major AI provider that has published an independently reviewable per-prompt energy and water methodology, released in August 2025. If a company only gives you a marketing claim with no methodology attached, treat the number as a lower bound, not a fact.
3. Look at where the data center serving you actually sits
Data centers in drought-prone regions (parts of Arizona, Texas, and the broader Southwest US) draw on water systems that are already under strain, while a data center in a wetter climate or one using closed-loop, air-based cooling has a much smaller water footprint for the same compute. You usually can't pick your provider's region, but you can factor this in when a company advertises "green" data centers — ask which cooling method, not just which power source.
4. Batch heavy-generation tasks instead of iterating live
Every regenerate-and-discard cycle on an image or video model repeats the full resource cost. In my testing, cutting live trial-and-error regeneration in favor of one detailed prompt up front cut my own image-generation requests by roughly half for the same final output.
5. Default to the smallest model that gets the job done
Reasoning-mode and largest-tier models cost meaningfully more energy per query than standard chat models, and for routine tasks — email drafts, summaries, basic code — the accuracy difference rarely justifies it. My best AI models comparison breaks down which tiers are overkill for common tasks.
Example prompts you can copy
These help you audit your own AI footprint or push a vendor for real numbers instead of marketing claims:
“ I use [AI tool] for [list your top 3 use cases]. Which of these tasks (text chat, image generation, video generation, reasoning/deep-think mode) carries the highest per-query energy and water cost, and why? “
“ Has [AI company] published an independently reviewable methodology for per-prompt energy or water use, similar to Google's August 2025 Gemini disclosure? If not, what's the most recent third-party estimate available? “
“ I want to cut the resource footprint of my AI usage without giving up capability. Suggest which of my regular tasks could move to a smaller or non-reasoning model with minimal quality loss. “
Common mistakes to avoid
The biggest mistake I see, including in my own early coverage of this topic, is quoting the 2024 Washington Post/UC Riverside figure (519 mL per email) as if it's still current — the same researcher's own 2025 revision is roughly 30x lower. Second, people compare "AI water use" to "data center water use" as if they're the same thing; AI is the fastest-growing slice, but plenty of data center water use predates the AI boom and isn't driven by it. Third, don't assume efficiency gains solve the problem: the report and Google's own disclosure both show massive per-query efficiency improvements (Google cites a 33x drop in energy per median prompt over 12 months), yet total footprint still rises, because usage volume is growing faster than efficiency. Fourth, don't treat "renewable-powered" as the same as "low-water" — a facility running on clean electricity can still draw heavily on a local aquifer for cooling.
Tools that make this easier
None of the tracking above requires a paid subscription, but a few things are worth pairing with it. If you're already watching what AI is costing your household, my breakdown of why AI subscriptions got so expensive covers the direct-cost side, while my map of how AI data centers are driving up power bills covers the electricity side of this same resource story, region by region. The financing behind why so much data center capacity is getting built at once, water and land footprint included, is also behind the private-power buildout I covered in AI fortunes reviving the private power debate. If you're picking tools for a small team and want to weigh capability against cost and footprint together, start with best AI tool for small business and AI tool ratings. And if you want the broader verification habits I used to separate the solid UN figures here from vaguer viral claims, that's covered in why AI mania is eviscerating good decision-making.
My take
The UN report is careful to say it's "not a case against AI," and I'd agree with that framing — the honest read is that AI's resource footprint is real, growing, and unevenly distributed, not that AI itself is a lost cause. What bothered me most in my research wasn't any single number, it was the concentration: over 90% of AI-specialized compute sits in two countries, while the water and land footprint of running that compute increasingly falls on wherever the data centers physically are, which is often not where the AI's users live. If you use AI tools daily, the actionable part is small but real: know which of your habits (image/video generation, reasoning mode) are the expensive ones, ask providers for real per-prompt numbers instead of marketing claims, and don't assume "efficient" and "sustainable at current growth rates" mean the same thing — right now, per the UN's own numbers, they don't.
Frequently Asked Questions
Is it true that AI could use as much water as 1.3 billion people by 2030?
That's the UN University's June 2026 projection: AI-related data centers' water footprint by 2030 would equal the basic annual domestic water needs of the 1.3 billion people living in Sub-Saharan Africa. It's a projection based on current buildout trends, not a guaranteed outcome — efficiency gains or policy changes could shift it either way.
How much water does a single ChatGPT or Gemini prompt actually use?
Estimates vary widely by methodology. Google's own audited disclosure puts a median Gemini text prompt at 0.26 mL of water. OpenAI has cited roughly 0.32 mL per average ChatGPT query. Older 2024 estimates as high as 519 mL per email are now considered outdated by the same researchers who produced them.
Does using a smaller or non-reasoning AI model actually reduce resource use?
Yes, meaningfully. The UN report found a single AI-generated image uses about 1,450 times the energy of a basic text classification task, and a video roughly 200,000 times as much. Sticking to standard chat mode instead of reasoning mode for routine tasks is one of the few resource-saving choices an individual user can directly control.
Which AI companies publish real per-prompt energy and water numbers?
As of this writing, Google is the only major provider with a published, independently reviewable per-prompt methodology, released in August 2025. Other companies' figures, including OpenAI's, are self-reported without full public methodology.
Is this an argument for not using AI tools at all?
No — the report's own authors call it "a call for using it responsibly," not a case against AI. The practical takeaway is choosing the right tool and mode for the task, not avoiding AI entirely.